Convegno Calcolo ad Alte Prestazioni "Biocomputing" Bio-molecular diagnosis through Random Subspace Ensembles of Learning Machines Alberto Bertoni, Raffaella Folgieri, Giorgio Valentini DSI Dipartimento di Scienze dell Informazione Università degli Studi di Milano {bertoni,folgieri,valentini}@dsi.unimi.it http://homes.dsi.unimi.it/~valenti
Outline Bio-molecular diagnosis of tumors using machine learning methods Current approaches to automatic bio-molecular diagnosis Random Subspace (RS) ensemble: experimental results on a case study Combining feature selection and RS ensemble On-going work: RP-ensembles
Bio-molecular diagnosis of malignancies: motivations Traditional clinical diagnostic approaches may sometimes fail in detecting tumors (Alizadeh et al. 2001) Several results showed that bio-molecular analysis of malignancies may help to better characterize malignancies (e.g. gene expression profiling) Information for supporting both diagnosis and prognosis of malignancies at bio-molecular level may be obtained from high-throughput biotechnologies (e.g. DNA microarray)
Bio-molecular diagnosis of malignancies: current approaches Huge amount of data available from biotechnologies: analysis and extraction of significant biological knowledge is critical Current approaches: statistical methods and machine learning methods (Golub et al., 1999; Furey et al., 2000; Ramaswamy et al., 2001; Dudoit et al. 2002; Lee & Lee, 2003; Weston et al., 2003, Dettling et al., 2003, Dettling 2004, Zhou et al, 2005, Zhang et al., 2006).
Main problems with gene expression data for bio-molecular diagnosis High dimensionality Low cardinality Curse of dimensionality Data are usually noisy: Gene expression measurements Labeling errors
Current approaches against the curse of dimensionality Selection of significant subsets of components (genes) e.g.: filter methods, forward selection, backward selection, recursive feature elimination, entropy and mutual information based feature selection methods (see Guyon & Ellisseef, 2003 for a recent review). Extraction of significant subsets of features e.g.: Principal Component Analysis or Independent Component Analysis Anyway, both approaches have problems...
An alternative approach based on ensemble methods Random subspace (RS) ensembles: RS (Ho, 1998) reduce the high dimensionality of the data by randomly selecting subsets of genes. Aggregation of different base learners trained on different subsets of features may reduce variance and improve diversity D 1 h 1 D Algorithm Aggregation h D m h m
The RS algorithm Input: a d-dimensional labelled gene expression data set D - a learning algorithm L - subspace dimension n<d - number of the base learners I Output: - Final hypothesis h ran :X C computed by the ensemble begin for i = 1 to I begin D i = Subspace_projection(D,n) H i = L(D i ) end h ran (x)=argmax t C card({i h i (x)=t}) end
Reasons for applying RS ensembles to the bio-molecular diagnosis of tumors Gene expression data are usually very high dimensional, and RS ensembles reduce the dimensionality and are effective with high dimensional data (Skurichina and Duin, 2002) Co-regulated genes show correlated gene expression levels (Gasch and Eisen, 2002), and RS ensembles are effective with correlated sets of features (Bingham and Mannila, 2001) Random projections may improve the diversity between base learners Overall accuracy of the ensemble may be enhanced through aggregation techniques (at least w.r.t. the variance component of the error)
Colon adenocarcinoma diagnosis Data (Alon et al., 1999): 62 samples 40 colon tumors 22 normal colon samples 2000 genes Methods: RS ensembles with linear SVMs as base learners Single linear SVMs Software: C++ NEURObjects library Hardware: Avogadro cluster of Xeon double processor workstations (Arlandini, 2005)
Results Colon tumor prediction (5 fold cross validation)
Colon tumor prediction: error as a function of the susbspace dimension Single SVM test error
Average base learner error The better accuracy of the RS ensemble does not simply depend on the better accuracy of their component base learners
- Open problems with RS methods 1. Can we explain the effectiveness of RS through the diversity of the base learners? 2. Can we get a bias-variance interpretation? 3. What about the optimal subspace dimension? 4. Are feature selection and random subspace ensemble approaches alternative, or it may be useful to combine them?
Combining feature selection and random subspace ensemble methods Random Subspace on Selected Features (RS-SF algorithm) A two-steps algorithm: 1. Select a subset of features (genes) according to a suitable feature selection method 2. Apply the random subspace ensemble method to the subset of selected features
Results on combining feature selection with random subspace ensembles Colon data set (Alon, 1999) 5-fold cross validation
Comparison with other methods Methods Estimated error LogitBoost (Dettling and Buhlmann, 2003) Bagging (Valentini et al., 2004) BagBoost (Dettling, 2004) Random Forest (Breiman, 2001) Random Subspace SVM PAM (Tibshirani et al. 2002) DLDA (Dudoit et al. 2002) knn 0.1914 0.1286 0.1610 0.1486 0.0968 0.1129 0.1190 0.1286 0.1638 Colon data set: generalization error estimated through crossvalidation or multiple-hold out techniques
An on-going development: Supervised Randomly Projected Ensembles (RP-ensembles): Recent work on unsupervised analysis of complex bio-molecular data (Bertoni and Valentini, 2006) showed that random projections obeying the Johnson-Lindenstrauss lemma can be used for: Discovering structures in bio-molecular data Validating clustering results Improving clustering results Random projections to lower dimensional subspaces can be applied to supervised analysis (e.g. bio-molecular diagnosis)?
Conclusions RS ensembles can improve the accuracy of biomolecular diagnosis characterized by very high dimensional data They could be also easily applied to heterogeneous bio-molecular and clinical data. A new promising approach consists in combining state of the art feature (gene) selection methods and RS ensembles RS ensembles are computationally intensive but can be easily parallelized using clusters of workstations (e.g. in a MPI framework).